Re: [PATCH] tcp: fix ICMP-RTO war

[David Miller <davem@davemloft.net>]

From: Damian Lukowski <damian@tvk.rwth-aachen.de>
Date: Mon, 25 Jan 2010 16:07:43 +0100

> @@ -530,7 +530,11 @@ static inline void tcp_bound_rto(const struct sock *sk)
>  
>  static inline u32 __tcp_set_rto(const struct tcp_sock *tp)
>  {
> -	return (tp->srtt >> 3) + tp->rttvar;
> +	u32 rto = (tp->srtt >> 3) + tp->rttvar;
> +	if (unlikely(rto < TCP_RTO_MIN))
> +		return TCP_RTO_MIN;
> +	else
> +		return rto;
>  }

The min RTO is now a runtime variable, TCP_RTO_MIN is merely the
default, so we should use tcp_rto_min() for obtaining that value.

And if we make this change, we might want to delete the comment in
tcp_set_rto() which claims:

	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
	 * guarantees that rto is higher.
	 */
	tcp_bound_rto(sk);

And we have shown here at least one case where that is not true.
:-)

I've looked at Denys's traces and your analysis, and I still
can't figure out who the true culprit is that lets us get into
such a state that RTO is evaluated so low...